CN117172134B

CN117172134B - Moon surface multiscale DEM modeling method based on fusion terrain features

Info

Publication number: CN117172134B
Application number: CN202311352195.1A
Authority: CN
Inventors: 陈玉敏; 陈国栋; 陈若璇; 马鉴燊; 苏恒; 李璋晖
Original assignee: Wuhan University WHU
Current assignee: Wuhan University WHU
Priority date: 2023-10-19
Filing date: 2023-10-19
Publication date: 2024-01-16
Anticipated expiration: 2043-10-19
Also published as: CN117172134A

Abstract

The invention provides a lunar surface multiscale DEM modeling method and system based on a fusion terrain feature, which belong to the technical field of terrain mapping and comprise the following steps: acquiring the lunar surface south pole LOLA DEM data and the LROC NAC DEM data, and determining a plurality of high-resolution DEM data sets and a plurality of low-resolution DEM data sets; applying the residual fusion convolutional neural network to the DEM fine modeling to obtain a multi-level resolution DEM model; reconstructing the high-resolution DEM by combining the topographic features with a neural network, constructing a topographic feature related loss function, and reconstructing the moon south pole high-precision large-range fine DEM with any scale. According to the invention, a deep learning method is adopted to manufacture a training sample set, fine modeling is carried out step by step on moon south pole DEM data, a residual fusion neural network is utilized to effectively extract and integrate spatial features, super-resolution reconstruction of any scale is supported, and a new idea is provided for fine modeling of moon south pole DEM.

Description

Moon surface multiscale DEM modeling method based on fusion terrain features

Technical Field

The invention relates to the technical field of topographic mapping, in particular to a lunar surface multiscale DEM modeling method and system based on converged topographic features.

Background

With the continuous development of space exploration technology, lunar and deep space exploration has become the leading edge and challenge of space information science. By utilizing the remote sensing observation data of the lunar orbit, the digital elevation model (Digital Elevation Model, DEM) products of the lunar topography with different resolutions manufactured by each research institution provide powerful support and guarantee for engineering detection and scientific research such as lunar surface topography feature analysis, landing zone evaluation and site selection, patrol machine path planning and the like. The moon south pole topography is complex, the elevation fluctuation is large, the oldest and deepest impact pit is provided, the engineering difficulty of soft landing detection is higher than that of a middle-low latitude area, the south pole fine earth surface topography data can not be detected and researched, and the high-precision DEM product provides a vital basic data support for a large engineering task.

Aiming at the remote sensing observation data of the lunar orbit device, each research institution makes digital elevation model products of the topography of the whole moon or partial moon with different resolutions. For example, moon area topography products with different resolutions of 7, 20 and 50 meters are manufactured based on goddess Chang E second stereoscopic images; a DEM with a resolution of 10 meters in whole month is manufactured by using a SELENE stereoscopic image in Japan and a laser altimeter; different resolutions DEM of 5, 10, 20, 30 and 60 meters covering different latitudes are manufactured by adopting laser altimeter (Lunar Orbiter Laser Altimeter, LOLA) data, wherein the resolution of the DEM covering the range from 87.5 DEG to the polar point reaches 5 meters, and the accuracy is insufficient to meet the actual requirements of lunar exploration engineering. German aerospace center (DLR) fabricated 2m resolution DEM of the sandaton local area of south pole by photogrammetry using LRO (Lunar Reconnaissance Orbiter, LRO) narrow angle camera images (Narrow Angle Camera, NAC) to support the lunar south pole detection site selection needs of the european hollow office (DLR, 2012). The existing high-resolution DEM product manufactured based on LRO NAC can reach the resolution of a local area of 0.5-2 meters, but the coverage range of the high-resolution DEM product is very limited. However, the existing multi-scale DEM data have the problems of scale breaking, scale discontinuity, data redundancy and the like, and cannot meet the requirements of multi-scale topographic analysis, so that the development of large-area and even global high-resolution lunar surface topographic data is a key problem to be solved currently.

At present, three methods are mainly used for carrying out fine modeling on a moon south pole DEM, namely three-dimensional modeling based on a stereoscopic image, li Chunlai et al screen 628-orbit images of a satellite of the Change No. one, and the technical problems of lack of lunar surface textures, low albedo and low-contrast moon image matching are effectively solved based on a SIFT feature matching and least square matching combined method. Wang Fenfei et al utilize the higher-resolution goddess satellite image of resolution ratio, carry out pyramid image matching based on the RPD model, realize the pixel-by-pixel matching, then utilize the anti-distance interpolation method to produce the three-dimensional topography of moon table, fill up the leak based on radial basis function at last. The lunar surface morphology is complex, the traditional cavity filling method still cannot well recover the lunar surface details, and the cavity filling method still needs further research.

Secondly, based on three-dimensional modeling of laser point cloud, li Chunlai et al generate DEM data with spatial resolution of 3km by using the goddess E laser altimeter data, and the accuracy is superior to U.S. ULCN2005 and is equivalent to DEM data of a Japan SELENE laser altimeter. Hu Wenmin and the like are used for carrying out track intersection analysis and a parallel difference method research on a Change's first laser altimeter, and the disappearance or weakening of the DEM banding phenomenon generated on the parallel difference can obviously improve the quality of the DEM. Hu et al propose a local terrain constraint based cross adjustment method using a new lunar global digital elevation model derived from the laser altimeter data of the goddess E No. 1. Whereas DEMs generated based on laser altimeters have relatively no high spatial resolution.

Thirdly, three-dimensional modeling based on multi-source data fusion. Wu et al propose a new combined block adjustment method combining the image of the goddess Chang No. 2 in China with the LOLA data of the lunar survey orbit aircraft of the U.S. astronavigation agency for accurate lunar terrain modeling, which avoids the inconsistent phenomenon caused by the independent use of an image or a laser altimeter. Barker et al, via a lunar orbit vehicle laser altimeter and SELENE terrain camera, generate a new lunar terrain model, and by combining two co-registered datasets, a near global DEM with high geodetic accuracy is obtained without the need for surface interpolation. Li Xin et al construct three-dimensional terrain with 1.5 m resolution of a local key area of a south pole of a moon by using LOLA DEM data, narrow-angle camera images of a lunar survey orbit aircraft and the like, and the modeling result has better consistency with high-precision LOLA data.

However, the lunar surface DEM fine modeling method is complex in algorithm, and the generated high-resolution DEM is limited in accuracy because high-frequency information is not introduced. With the continued development of convolutional neural networks, many researchers have explored methods of applying convolutional neural networks to DEM fine-scale deduction. Chen et al utilize three-layer convolutional neural networks to perform feature extraction, feature compression and feature transformation on DEMs to achieve reconstruction of low-resolution DEMs to high-resolution DEMs. Xu et al propose a convolutional neural network training learning method based on transfer learning and gradient prior, and apply the convolutional neural network training learning method to two network structures, namely EDSR and SRCNN. Li et al integrate the topography knowledge of ridge lines and valleys into a deep learning framework, and fill up the DEM missing area better. However, the current research is focused on directly applying the deep convolutional neural network to the reconstruction of the DEM, the network structure design is not better optimized according to the characteristics of the DEM data, the importance influence of different terrain features is not fully considered, and meanwhile, the problems of fixed research area and poor migration performance still exist in most models.

Disclosure of Invention

The invention provides a lunar surface multiscale DEM modeling method and system based on a converged terrain feature, which are used for solving the defects that an algorithm aiming at lunar surface DEM fine modeling in the prior art is complex to implement, the precision is low and migration and transformation cannot be well realized.

In a first aspect, the invention provides a lunar surface multiscale DEM modeling method based on a fused terrain feature, comprising the following steps:

obtaining LOLA DEM data and LROC NAC DEM data of a moon surface south pole, and obtaining a plurality of preset high-resolution DEM data sets and a plurality of preset low-resolution DEM data sets based on the LOLA DEM data and the LROC NAC DEM data;

extracting topographic feature map data sets corresponding to the plurality of preset high-resolution DEM data by using a topographic feature extraction operator;

constructing a residual fusion convolutional neural network, and training the residual fusion convolutional neural network based on the plurality of preset high-resolution DEM data sets, the plurality of preset low-resolution DEM data sets and the topographic feature map data set to obtain a multi-level resolution DEM model;

obtaining a plurality of preset high-precision DEM data by adopting the multi-level resolution DEM model and a plurality of preset scale resolution DEM data;

determining a preset small-range DEM data set by the plurality of preset high-precision DEM data, and training the multi-level resolution DEM model based on the preset small-range DEM data set to obtain a preset small-range DEM super-resolution model;

And inputting the plurality of preset high-precision DEM data into the preset small-range DEM super-resolution model to obtain preset small-range fine DEM data.

According to the lunar surface multiscale DEM modeling method based on the fusion terrain features, LOLA DEM data and LROC NAC DEM data are obtained, a plurality of preset high-resolution DEM data sets and a plurality of preset low-resolution DEM data sets are obtained based on the LOLA DEM data and the LROC NAC DEM data, and the method comprises the following steps:

acquiring a first-level resolution DEM, a second-level resolution DEM, a third-level resolution DEM, a fourth-level length resolution DEM and a fifth-level resolution DEM in the LOLA DEM data in the same area range in the lunar surface south pole, and a sixth-level resolution DEM in the LROC NAC DEM data to obtain the plurality of preset high-resolution DEM data sets and the plurality of preset low-resolution DEM data sets, wherein the first-level resolution, the second-level resolution, the third-level resolution, the fourth-level resolution, the fifth-level resolution and the sixth-level resolution become higher in sequence;

projecting the plurality of preset low-resolution DEM data sets to a coordinate system of the plurality of preset high-resolution DEM data sets to obtain a plurality of preset low-resolution DEM data sets and a plurality of preset high-resolution DEM data sets of a unified coordinate system;

Cutting a plurality of preset high-resolution DEM data sets of a unified coordinate system by adopting a first preset pixel square, and removing a region of which the boundary is smaller than the first preset pixel square to obtain a plurality of preset high-resolution DEM data sets of which the first integer times is larger than the first preset pixel square;

cutting the plurality of preset low-resolution DEM data sets of the unified coordinate system by using boundary range lines obtained by the plurality of preset high-resolution DEM data sets of the first integer multiple and a second preset pixel square to obtain a plurality of preset low-resolution DEM data sets of the second integer multiple;

wherein the first integer multiple is twice the second integer multiple.

According to the lunar surface multiscale DEM modeling method based on the fusion of the topographic features, provided by the invention, the topographic feature map dataset corresponding to the plurality of preset high-resolution DEM data is extracted by using the topographic feature extraction operator, and the method comprises the following steps:

extracting ridge line data sets and valley line data sets corresponding to the plurality of preset high-resolution DEM data by adopting the topographic feature extraction operator;

converting the ridge line data set and the valley line data set into raster data, reclassifying the raster data and giving an identification terrain feature line, cutting a preset low-resolution gradient map by utilizing boundary range lines of the plurality of preset high-resolution DEM data sets, and dividing the preset low-resolution gradient map into a plurality of high-resolution terrain feature line map pixel blocks;

And the plurality of high-resolution topographic feature map pixel blocks are corresponding to the plurality of preset high-resolution DEM data, so that the topographic feature map data set is obtained.

According to the lunar surface multiscale DEM modeling method based on the fusion topography features, provided by the invention, a residual fusion convolutional neural network is constructed, and the method comprises the following steps:

the residual fusion convolutional neural network comprises a feature shallow layer extraction module, a middle layer module and a space residual attention enhancement module which are connected in sequence;

the feature shallow layer extraction module comprises a convolution layer and a convolution kernel with the size of 3 multiplied by 3;

the middle layer module comprises 30 residual error fusion modules, and each residual error fusion module comprises four residual error modules;

the enhanced spatial residual attention module comprises a convolution group consisting of a 1 multiplied by 1 convolution layer, a cavity convolution layer and three 3 multiplied by 3 convolution layers, a characteristic splicing layer, the 1 multiplied by 1 convolution layer and a Sigmoid layer which are connected in sequence.

According to the lunar surface multiscale DEM modeling method based on the fusion topographic features, provided by the invention, the residual fusion convolutional neural network is trained based on the preset high-resolution DEM data sets, the preset low-resolution DEM data sets and the topographic feature map data sets to obtain a multistage resolution DEM model, and the method comprises the following steps:

Dividing a dataset formed by the plurality of preset high-resolution DEM datasets, the plurality of preset low-resolution DEM datasets and the topographic feature map dataset into a training dataset and a testing dataset;

and initializing network parameters by using the training data set, and integrating the topographic feature map data set into the total loss function of the residual fusion convolutional neural network to obtain the multi-resolution DEM model.

According to the lunar surface multiscale DEM modeling method based on the fusion topographic features, which is provided by the invention, a topographic feature map data set is fused into the overall loss function of the residual fusion convolutional neural network, and the method comprises the following steps:

the global loss function is obtained by the number of pixels, the real elevation value and the predicted elevation value contained in the single DEM;

obtaining an X-direction gradient according to an X-direction elevation value and a pixel size, obtaining a Y-direction gradient according to a Y-direction elevation value and a pixel size, obtaining a gradient according to the X-direction gradient, the Y-direction gradient and a default tolerance, and obtaining the gradient loss function according to a real gradient value of any DEM unit cell in the gradient and a predicted gradient value of any DEM unit cell;

obtaining the topographic feature line loss function by the corresponding class identifier of each DEM pixel value, the corresponding pixel of the topographic feature line of the topographic feature point in the single DEM, the real elevation value and the predicted elevation value;

The global loss function is formed by a first adjustment weight, the gradient loss function, a second adjustment weight, the topographic feature line loss function and the global loss function.

According to the lunar surface multiscale DEM modeling method based on the fusion terrain features, provided by the invention, the multilevel resolution DEM model and a plurality of preset scale resolution DEM data are adopted to obtain a plurality of preset high-precision DEM data, and the method comprises the following steps:

inputting the first length resolution DEM and the second length resolution DEM into the plurality of preset scale resolution DEM data to obtain a reconstructed second length resolution DEM;

inputting the reconstructed second length resolution DEM and third length resolution DEM into the plurality of preset scale resolution DEM data to obtain a reconstructed third length resolution DEM;

inputting the reconstructed third length resolution DEM and fourth length resolution DEM into the plurality of preset scale resolution DEM data to obtain a reconstructed fourth length resolution DEM;

and inputting the reconstructed fourth length resolution DEM and the reconstructed fifth length resolution DEM into the plurality of preset scale resolution DEM data to obtain the reconstructed fifth length resolution DEM.

According to the lunar surface multiscale DEM modeling method based on the fusion terrain features, a preset small-range DEM data set is determined by the preset high-precision DEM data, the multistage resolution DEM model is trained based on the preset small-range DEM data set, and a preset small-range DEM super-resolution model is obtained, and the method comprises the following steps:

Constructing a sixth length resolution DEM from said reconstructed fifth length resolution DEM;

and training the multi-level resolution DEM model based on the sixth length resolution DEM to obtain a DEM super-resolution model from the fifth length resolution to the sixth length resolution.

According to the lunar surface multiscale DEM modeling method based on the fusion topographic features, the preset high-precision DEM data are input into the preset small-range DEM super-resolution model to obtain the preset small-range fine DEM data, and the method comprises the following steps:

and inputting the reconstructed DEM with the fifth length resolution into a DEM super-resolution model from the fifth length resolution to the sixth length resolution to obtain sixth length resolution fine DEM data.

According to the lunar surface multiscale DEM modeling method based on the fusion topographic features, the method inputs the preset high-precision DEM data into the preset small-range DEM super-resolution model to obtain the preset small-range fine DEM data, and then the method further comprises the following steps:

and evaluating the super-resolution model precision of the pre-set small-range DEM super-resolution model.

In a second aspect, the present invention also provides a lunar surface multiscale DEM modeling system based on fused topographical features, comprising:

The acquisition module is used for acquiring LOLA DEM data and LROC NAC DEM data of the south pole of the lunar surface, and acquiring a plurality of preset high-resolution DEM data sets and a plurality of preset low-resolution DEM data sets based on the LOLA DEM data and the LROC NAC DEM data;

the extraction module is used for extracting topographic feature map data sets corresponding to the plurality of preset high-resolution DEM data by using a topographic feature extraction operator;

the first training module is used for constructing a residual fusion convolutional neural network, and training the residual fusion convolutional neural network based on the plurality of preset high-resolution DEM data sets, the plurality of preset low-resolution DEM data sets and the topographic feature map data set to obtain a multi-level resolution DEM model;

the fine module is used for obtaining a plurality of preset high-precision DEM data by adopting the multi-level resolution DEM model and a plurality of preset scale resolution DEM data;

the second training module is used for determining a preset small-range DEM data set from the plurality of preset high-precision DEM data, and training the multi-level resolution DEM model based on the preset small-range DEM data set to obtain a preset small-range DEM super-resolution model;

and the processing module is used for inputting the plurality of preset high-precision DEM data into the preset small-range DEM super-resolution model to obtain preset small-range fine DEM data.

In a third aspect, the invention also provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing a lunar surface multiscale DEM modeling method based on fused topographical features as described in any one of the above when executing the program.

In a fourth aspect, the present invention also provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements a method of multi-scale DEM modeling of a lunar surface based on fused topographical features as described in any of the above.

According to the lunar surface multiscale DEM modeling method and system based on the fusion topography features, training sample set manufacturing is conducted through a deep learning method, fine modeling is conducted step by step on lunar antarctic DEM data, space features are effectively extracted and integrated through a residual fusion neural network, reconstruction of super-resolution of any scale is supported, and a new idea is provided for fine modeling of lunar antarctic DEM.

Drawings

In order to more clearly illustrate the invention or the technical solutions of the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described, and it is obvious that the drawings in the description below are some embodiments of the invention, and other drawings can be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a schematic flow diagram of a lunar surface multiscale DEM modeling method based on a fused topography feature;

FIG. 2 is a second flow chart of a lunar surface multiscale DEM modeling method based on the fusion of topographic features;

FIG. 3 is a network structure diagram of a residual fusion convolutional neural network provided by the invention;

FIG. 4 is a block diagram of a residual fusion module provided by the present invention;

FIG. 5 is a block diagram of an enhanced spatial residual attention module provided by the present invention;

FIG. 6 is a schematic structural diagram of a lunar surface multi-scale DEM modeling system based on fused topographical features provided by the present invention;

fig. 7 is a schematic structural diagram of an electronic device provided by the present invention.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings, and it is apparent that the described embodiments are some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

Fig. 1 is one of flow diagrams of a lunar surface multiscale DEM modeling method based on a fused terrain feature according to an embodiment of the invention, as shown in fig. 1, including:

step 100: obtaining LOLA DEM data and LROC NAC DEM data of a moon surface south pole, and obtaining a plurality of preset high-resolution DEM data sets and a plurality of preset low-resolution DEM data sets based on the LOLA DEM data and the LROC NAC DEM data;

step 200: extracting topographic feature map data sets corresponding to the plurality of preset high-resolution DEM data by using a topographic feature extraction operator;

step 300: constructing a residual fusion convolutional neural network, and training the residual fusion convolutional neural network based on the plurality of preset high-resolution DEM data sets, the plurality of preset low-resolution DEM data sets and the topographic feature map data set to obtain a multi-level resolution DEM model;

step 400: obtaining a plurality of preset high-precision DEM data by adopting the multi-level resolution DEM model and a plurality of preset scale resolution DEM data;

step 500: determining a preset small-range DEM data set by the plurality of preset high-precision DEM data, and training the multi-level resolution DEM model based on the preset small-range DEM data set to obtain a preset small-range DEM super-resolution model;

Step 600: and inputting the plurality of preset high-precision DEM data into the preset small-range DEM super-resolution model to obtain preset small-range fine DEM data.

As shown in fig. 2, first, lunar antarctic LOLA DEM data and LROC NAC DEM data are acquired from a LOLA official website, and coordinate transformation and cutting preprocessing are performed to obtain a plurality of corresponding high-resolution DEM and low-resolution DEM data sets. In addition, the ridge line data set and the valley line data set corresponding to the high-resolution DEM data are extracted and constructed by utilizing the topographic feature extraction operator

And then constructing a residual fusion convolutional neural network combining with terrain features for modeling any fine scale of the lunar surface, dividing the obtained data set into a training data set and a test data set, respectively training the constructed residual fusion convolutional neural network by using high-resolution DEM data and low-resolution DEM data corresponding to each training data set and high-resolution terrain feature images, initializing network parameters, and integrating the gradient and the extracted terrain features into a loss function of the neural network to obtain a multi-resolution DEM model corresponding to each level of resolution DEM.

Obtaining finer-scale high-precision DEM data by adopting a multi-level resolution DEM model obtained through training and DEM data with multiple scale resolutions; and constructing and training a network by using the obtained finer-scale high-precision DEM data, initializing network parameters, and obtaining a preset small-range DEM super-resolution model.

And finally, inputting the plurality of high-precision DEM data into a preset small-range DEM super-resolution model to obtain preset small-range fine DEM data.

According to the invention, a deep learning method is adopted to manufacture a training sample set, fine modeling is carried out step by step on moon south pole DEM data, a residual fusion neural network is utilized to effectively extract and integrate space characteristics, super-resolution reconstruction of any scale is supported, and a new idea is provided for fine modeling of moon south pole DEM.

Based on the above embodiment, obtaining the LOLA DEM data and the LROC NAC DEM data, obtaining a plurality of preset high resolution DEM data sets and a plurality of preset low resolution DEM data sets based on the LOLA DEM data and the LROC NAC DEM data, including:

acquiring a first-level resolution DEM, a second-level resolution DEM, a third-level resolution DEM, a fourth-level resolution DEM and a fifth-level resolution DEM in the LOLA DEM data in the same area range in the south pole of the lunar surface, and a sixth-level resolution DEM in the LROC NAC DEM data to obtain a plurality of preset high-resolution DEM data sets and a plurality of preset low-resolution DEM data sets, wherein the first-level resolution, the second-level resolution, the third-level resolution, the fourth-level resolution, the fifth-level resolution and the sixth-level resolution become higher in sequence;

wherein the first integer multiple is twice the second integer multiple.

Specifically, the embodiment of the invention carries out corresponding coordinate transformation on LOLA 60 m resolution DEM (60 DEG S-90 DEG S), 30 m resolution DEM (75 DEG S-90 DEG S), 20 m resolution DEM (80 DEG S-90 DEG S), 10 m resolution DEM (85 DEG S-90 DEG S) and 5 m resolution DEM (87.5 DEG S-90 DEG S) in the same area range of the south pole of the lunar surface, and 2 m resolution DEM distributed on the south pole of the lunar in LROC NAC data, and projects the low resolution DEM to the same coordinate system as the high resolution DEM.

Selecting an area with obvious elevation value change in the high-resolution DEM, cutting by using a block with the size of 192 multiplied by 192 pixels, and discarding the area with the size of less than 192 multiplied by 192 pixels at the boundary to finally obtain the DEM with the integer multiple of 192 multiplied by 192; and simultaneously, cutting the low-resolution DEM by using the obtained boundary range line, dividing the cut low-resolution DEM into a plurality of 96 multiplied by 96 pixel blocks, firstly establishing a corresponding relation between the high-resolution DEM pixel blocks and the low-resolution DEM pixel blocks according to the resolution and the coverage range of the DEM products at each level, and generating four high-resolution DEM data sets of 60m-30m, 40m-20m, 20m-10m and 10m-5m, wherein the generated 30m is resampled to 40m to meet the 2 times super-resolution relation.

Based on the above embodiment, extracting the topographic feature map dataset corresponding to the plurality of preset high resolution DEM data using the topographic feature extraction operator includes:

Specifically, the embodiment of the invention utilizes a topographic feature extraction operator to extract and construct a ridge line and valley line data set corresponding to high-resolution DEM data, converts the ridge line and valley line data set into raster data, reclassifies the topographic feature lines to identify the topographic feature lines, cuts a low-resolution gradient map by using the boundary range lines obtained in the embodiment, divides the low-resolution gradient map into a plurality of 192 multiplied by 192 pixel blocks, establishes the corresponding relation between the high-resolution topographic feature map pixel blocks and the high-resolution DEM pixel blocks, and adds the divided topographic feature map into the data set.

Based on the above embodiment, constructing a residual fusion convolutional neural network includes:

Specifically, the embodiment of the invention constructs the residual fusion convolutional neural network of the combined terrain features modeling any fine scale of the lunar surface, the structure of the residual fusion convolutional neural network is shown in fig. 3, and the network comprises 3 modules. The first module is a feature shallow extraction module, which comprises a convolution layer, the convolution kernel size is 3×3, the input is a low resolution DEM, and the output is an extracted feature map with the size of 256. The second middle layer module is a combination of 30 Residual fusion modules, each Residual fusion module containing 4 Residual structures, and the Residual fusion module (Residual Feature fusion Module, RFM) contains four Residual Blocks (Residual Blocks), as shown in fig. 4, where the first three Residual Blocks can acquire local features added by Residual features. The residual features of the first three residual blocks are connected to the output of the last residual block as input to a 1 x 1 convolutional layer for fusing residual features and reducing spatial dimensions, and then combined with low-level features extracted from the previous RFM by element-wise addition. RFM can propagate the effects of intermediate features at a lower cost than simple stackable residual blocks, resulting in efficient feature extraction from DEM images and powerful representation capabilities. Unlike a simple stacking of 30 residual feature aggregation modules derived from RFMs, the framework of embodiments of the present invention also considers fusion of local features extracted from each RFM, considering each 10 RFMs as a base unit, which acts as a residual block in the RFM and may be referred to as "fused residual feature aggregation".

Further, the embodiment of the invention also provides an enhanced residual error module to maximize the effect of RFM in DEM fine modeling, so we add an Enhanced Spatial Residual Attention Module (ESRAM) at the end of RFM, as shown in fig. 5. The reduction in channel dimensions is achieved by the first 1 x 1 convolutional layer in the esam. A hole convolution layer is then designed to expand the receptive field of the block without changing the spatial dimensions. Hole convolution is widely used in the field of computer vision to obtain an expansion of receptive fields; thus, the dilation parameter of the convolutional layer is set to 6. The fact that the receptive field is not enlarged by setting larger hole parameters (e.g. 12) is because the extra 0-valued padding added to the edge of the DEM image occupies most of the feature map for convolution, resulting in weaker representation of the DEM image. The features are then reassembled from the set of three 3 x 3 convolutional layers and the low-level features are forwarded and concentrated to the input of one 1 x 1 convolutional layer using a jump connection to recover the channel dimension. Finally, a sigmoid layer is designed that generates a pixel-by-pixel attention factor that is multiplied by residual features extracted from the previously connected residual block.

For an up-sampling module with any scale, the method adopted is that the coordinates and the relative distance of each pixel in the LR space are calculated for each pixel in the high-resolution space, then the pixels are spliced and then sent into two full-connection layers for feature extraction, and the obtained features are used for predicting weights and offsets; secondly, the predicted weights are used for combining two groups of expert knowledge to obtain paired filter parameters; finally, the resulting filter parameters are used to convolve the input features to produce output features, thereby obtaining a feature map of the desired scale, and finally outputting the DEM at the desired resolution.

According to the invention, the information of different residual modules is fully fused by utilizing the residual fusion structure, the spatial attention module is introduced as a part of the residual modules, the spatial features are effectively extracted and integrated, the precision of DEM fine modeling is improved, and the super-resolution reconstruction of any scale is supported.

Based on the above embodiment, training the residual fusion convolutional neural network based on the plurality of preset high-resolution DEM data sets, the plurality of preset low-resolution DEM data sets, and the topographic feature map data set to obtain a multi-resolution DEM model, including:

The integrating the topographic feature map data set into the overall loss function of the residual fusion convolutional neural network comprises the following steps:

Specifically, the embodiment of the invention divides the obtained data set into a training data set and a test data set, respectively trains the constructed residual fusion convolutional neural network by using the high-resolution DEM data, the low-resolution DEM data and the high-resolution topographic feature images corresponding to the training data sets, initializes network parameters, and fuses the gradient and the extracted topographic features into a loss function of the neural network to obtain a training convergence model corresponding to each grade of resolution DEM.

For the super parameter settings in the model, 30 RFMs were used, with each RFM module containing four proposed esam blocks. The reduction factor of the 1 x 1 convolutional layer of each esam block is set to four. At the end, the number of channels of the output profile is set to 64, except for other convolution filters in the ESRAM. Optimization using Adam optimizer at counter-propagation, where β ₁ Set to 0.9, beta ₂ Set to 0.999, epsilon set to 10 ^-8 . All models were trained on a PyTorch platform, batch-wiseThe size (batch size) was set to 16 and training of 1000 epochs was performed using a single GeForce GTX 3090. The high resolution DEM image is randomly cropped to 96 x 96 pixels DEM, while the low resolution DEM image is cropped according to the corresponding scale. Data enhancement is performed using horizontal and vertical flipping. The initial learning rate is set to 1e-4 and adjusted using a learning rate strategy (cosine annealing LR) named cosine annealing learning rate. After 300 training sessions, MAE loss of the terrain line was considered as part of the overall function.

The loss function of the network training comprises a global loss part and a topographic feature loss part, and the calculation formula is as follows:

in the method, in the process of the invention,representing global losses, the terrain feature loss part comprises slope losses and terrain feature line losses, +.>Representing grade loss->Representing topographical feature line loss, < >>And->Representing the weights adjusted for the different loss sections. Wherein the global penalty is as follows:

in the middle ofRepresenting the true elevation value,/>Representing the predicted elevation value,/->Representing the number of pixels contained in a DEM.

And for gradient loss, calculating the gradient of the high-resolution DEM after coordinate transformation and the reconstructed DEM generated by a residual error fusion network by using a gradient extraction operator; the equation for the calculation of the gradient loss is as follows:

wherein:indicating gradient->And->Is the elevation value in the X direction, +.>And->For the elevation value in the Y direction, +.>Gradient in X direction>Is the gradient of Y direction>For default tolerance, the situation that gradient explosion occurs in the neural network back propagation derivative is prevented, and s represents the size of the pixel and the resolution. />Gradient value representing the ith DEM cell,/->Representing the predicted grade value of the ith DEM cell.

For the loss of the topographic feature line, we extract the topographic feature line (ridge line and valley line) from the high-resolution DEM by using the topographic feature line extracting operator, cut the high-resolution slope map by using the boundary range line obtained in the step 1.2, divide the map into a plurality of pixel blocks, establish the corresponding relation between the pixel blocks of the high-resolution topographic feature line and the pixel blocks of the high-resolution DEM, add the divided topographic feature map into the data set, and the topographic feature line loss calculation formula is as follows:

Wherein M represents pixels corresponding to topographic feature lines of topographic feature points in a DEM image,representing the class to which each DEM pixel value corresponds, the selectable value is 0 or 1,0 representing a non-topographical feature line pixel, and 1 representing a topographical feature line pixel.

In the network training process, the gradient and the loss part of the terrain characteristic line are introduced into the loss function, the constraint training is carried out on the network, and the problem that the consistency of the terrain characteristic of the neural network reconstructed DEM is not good under the premise of ensuring the accuracy is solved.

Based on the above embodiment, the obtaining a plurality of preset high-precision DEM data by using the multi-level resolution DEM model and a plurality of preset scale resolution DEM data includes:

Specifically, the embodiment of the invention starts from 60 m resolution DEM with the widest range of the south pole of the covered moon, gradually upwards restores finer DEM, firstly, 30 m resolution DEM of 60-90 degrees S of the south pole of the moon is reconstructed by utilizing 30 m resolution LOLA DEM and 60 m resolution LOLA DEM and residual fusion convolutional neural network; then, reconstructing the 20 m resolution DEM of the south pole of the whole moon through a residual fusion convolutional neural network by using the 20 m resolution LOLA DEM and the reconstructed 30 m resolution DEM; then, utilizing the LOLA DEM with the resolution of 10 meters and the reconstructed DEM with the resolution of 20 meters to construct the DEM with the resolution of 10 meters of the south pole of the whole moon; and finally, constructing the 5-meter resolution DEM of the whole moon south pole by using the 5-meter resolution LOLA DEM and the reconstructed 10-meter resolution DEM, and filling 5-meter high-precision DEM data which are lack in a 60-87.5-degree S range.

Based on the above embodiment, determining a preset small-range DEM data set from the plurality of preset high-precision DEM data, training the multi-level resolution DEM model based on the preset small-range DEM data set, to obtain a preset small-range DEM super-resolution model, including:

Specifically, according to the embodiment of the invention, training and manufacturing of a data set are carried out again according to the obtained 5-meter high-precision moon south pole DEM data in the 60-90-degree S range, and a DEM data set from 5 meters to 2 meters is constructed;

inputting the low-resolution DEM in the test data set into a trained DEM super-resolution model to obtain corresponding prediction high-resolution DEM small blocks, then splicing the small blocks to obtain a large-range prediction high-resolution DEM, and finally replacing the generated prediction DEM with the DEM data in a partial region to maximally maintain the modeling precision of the DEM.

Further, the newly constructed data set is subjected to network construction and training, network parameters are initialized, and a DEM super-resolution model corresponding to 5 meters to 2 meters is obtained.

Based on the above embodiment, inputting the plurality of preset high-precision DEM data into the preset small-range DEM super-resolution model to obtain preset small-range fine DEM data includes:

Specifically, according to the embodiment of the invention, the obtained 5-meter high-precision moon south pole DEM data in the 60-90 DEG S range is input into a training convergence DEM super-resolution model from 5 meters to 2 meters, and the 2-meter high-precision moon south pole DEM data in the 60-90 DEG S range of the moon south pole is obtained.

Based on the above embodiment, inputting the plurality of preset high-precision DEM data into the preset small-range DEM super-resolution model to obtain preset small-range fine DEM data, the method further includes:

Specifically, the embodiment of the invention trains DEM at different resolutionsEvaluating the precision of the super-resolution model obtained by training, predicting a prediction set by a model after training and convergence, comparing with an actual high-resolution DEM in a test data set, and calculating average absolute error (MAE), root Mean Square Error (RMSE) and maximum error (Km)) Average absolute gradient error (+)>) And the mean absolute error of the topographical features ( >) As an evaluation index; the calculation formula is as follows:

where N is the number of pixels in the high resolution DEM,refers to the elevation value in the actual high resolution DEM,/->Refers to predicting the value of a pixel in a high resolution DEM,/->Representing the corresponding class of each DEM pixel value, wherein the optional value is 0 or 1,0 represents the non-topographic feature line pixel, 1 represents the topographic feature line pixel, and +.>Gradient value representing the ith DEM cell,/->Representing the predicted grade value of the ith DEM cell.

The lunar surface multi-scale DEM modeling system based on the fusion topographic features provided by the invention is described below, and the lunar surface multi-scale DEM modeling system based on the fusion topographic features described below and the lunar surface multi-scale DEM modeling method based on the fusion topographic features described above can be correspondingly referred to each other.

Fig. 6 is a schematic structural diagram of a lunar surface multi-scale DEM modeling system based on fused terrain features according to an embodiment of the invention, as shown in fig. 6, including: an acquisition module 61, an extraction module 62, a first training module 63, a refinement module 64, a second training module 65, and a processing module 66, wherein:

the acquiring module 61 is configured to acquire LOLA DEM data and LROC NAC DEM data of a lunar surface south pole, and acquire a plurality of preset high-resolution DEM data sets and a plurality of preset low-resolution DEM data sets based on the LOLA DEM data and the LROC NAC DEM data; the extracting module 62 is configured to extract a topographic feature map dataset corresponding to the plurality of preset high resolution DEM data by using a topographic feature extracting operator; the first training module 63 is configured to construct a residual fusion convolutional neural network, train the residual fusion convolutional neural network based on the multiple preset high-resolution DEM data sets, the multiple preset low-resolution DEM data sets and the topographic feature map data sets, and obtain a multi-resolution DEM model; the fine module 64 is configured to obtain a plurality of preset high-precision DEM data by using the multi-level resolution DEM model and a plurality of preset scale resolution DEM data; the second training module 65 is configured to determine a preset small-range DEM data set from the multiple preset high-precision DEM data, and train the multi-level resolution DEM model based on the preset small-range DEM data set to obtain a preset small-range DEM super-resolution model; the processing module 66 is configured to input the multiple preset high-precision DEM data into the preset small-range DEM super-resolution model to obtain preset small-range fine DEM data.

Fig. 7 illustrates a physical schematic diagram of an electronic device, as shown in fig. 7, which may include: processor 710, communication interface (Communications Interface) 720, memory 730, and communication bus 740, wherein processor 710, communication interface 720, memory 730 communicate with each other via communication bus 740. Processor 710 may invoke logic instructions in memory 730 to perform a lunar surface multiscale DEM modeling method based on fused terrain features, the method comprising: obtaining LOLA DEM data and LROC NAC DEM data of a moon surface south pole, and obtaining a plurality of preset high-resolution DEM data sets and a plurality of preset low-resolution DEM data sets based on the LOLA DEM data and the LROC NAC DEM data; extracting topographic feature map data sets corresponding to the plurality of preset high-resolution DEM data by using a topographic feature extraction operator; constructing a residual fusion convolutional neural network, and training the residual fusion convolutional neural network based on the plurality of preset high-resolution DEM data sets, the plurality of preset low-resolution DEM data sets and the topographic feature map data set to obtain a multi-level resolution DEM model; obtaining a plurality of preset high-precision DEM data by adopting the multi-level resolution DEM model and a plurality of preset scale resolution DEM data; determining a preset small-range DEM data set by the plurality of preset high-precision DEM data, and training the multi-level resolution DEM model based on the preset small-range DEM data set to obtain a preset small-range DEM super-resolution model; and inputting the plurality of preset high-precision DEM data into the preset small-range DEM super-resolution model to obtain preset small-range fine DEM data.

Further, the logic instructions in the memory 730 described above may be implemented in the form of software functional units and may be stored in a computer readable storage medium when sold or used as a stand alone product. Based on this understanding, the technical solution of the present invention may be embodied essentially or in a part contributing to the prior art or in a part of the technical solution, in the form of a software product stored in a storage medium, comprising several instructions for causing a computer device (which may be a personal computer, a server, a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.

In another aspect, the present invention also provides a non-transitory computer readable storage medium having stored thereon a computer program which when executed by a processor is implemented to perform a method of multi-scale DEM modeling of a lunar surface based on fused topographical features provided by the methods described above, the method comprising: obtaining LOLA DEM data and LROC NAC DEM data of a moon surface south pole, and obtaining a plurality of preset high-resolution DEM data sets and a plurality of preset low-resolution DEM data sets based on the LOLA DEM data and the LROC NAC DEM data; extracting topographic feature map data sets corresponding to the plurality of preset high-resolution DEM data by using a topographic feature extraction operator; constructing a residual fusion convolutional neural network, and training the residual fusion convolutional neural network based on the plurality of preset high-resolution DEM data sets, the plurality of preset low-resolution DEM data sets and the topographic feature map data set to obtain a multi-level resolution DEM model; obtaining a plurality of preset high-precision DEM data by adopting the multi-level resolution DEM model and a plurality of preset scale resolution DEM data; determining a preset small-range DEM data set by the plurality of preset high-precision DEM data, and training the multi-level resolution DEM model based on the preset small-range DEM data set to obtain a preset small-range DEM super-resolution model; and inputting the plurality of preset high-precision DEM data into the preset small-range DEM super-resolution model to obtain preset small-range fine DEM data.

The apparatus embodiments described above are merely illustrative, wherein the elements illustrated as separate elements may or may not be physically separate, and the elements shown as elements may or may not be physical elements, may be located in one place, or may be distributed over a plurality of network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art will understand and implement the present invention without undue burden.

From the above description of the embodiments, it will be apparent to those skilled in the art that the embodiments may be implemented by means of software plus necessary general hardware platforms, or of course may be implemented by means of hardware. Based on this understanding, the foregoing technical solution may be embodied essentially or in a part contributing to the prior art in the form of a software product, which may be stored in a computer readable storage medium, such as ROM/RAM, a magnetic disk, an optical disk, etc., including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the method described in the respective embodiments or some parts of the embodiments.

Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and are not limiting; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A lunar surface multiscale DEM modeling method based on a fusion terrain feature is characterized by comprising the following steps:

acquiring lunar orbit aircraft laser altimeter digital elevation model LOLA DEM data of a lunar surface south pole and lunar survey orbit camera narrow angle camera digital elevation model LROC NAC DEM data, and acquiring a plurality of preset high-resolution DEM data sets and a plurality of preset low-resolution DEM data sets based on the LOLA DEM data and the LROC NAC DEM data;

inputting the plurality of preset high-precision DEM data into the preset small-range DEM super-resolution model to obtain preset small-range fine DEM data;

the multi-level resolution DEM model and a plurality of preset scale resolution DEM data are adopted to obtain a plurality of preset high-precision DEM data, and the method comprises the following steps:

Inputting the reconstructed fourth length resolution DEM and the reconstructed fifth length resolution DEM into the plurality of preset scale resolution DEM data to obtain a reconstructed fifth length resolution DEM;

determining a preset small-range DEM data set by the plurality of preset high-precision DEM data, training the multi-level resolution DEM model based on the preset small-range DEM data set to obtain a preset small-range DEM super-resolution model, wherein the method comprises the following steps of:

training the multi-level resolution DEM model based on the sixth length resolution DEM to obtain a DEM super-resolution model from the fifth length resolution to the sixth length resolution;

inputting the plurality of preset high-precision DEM data into the preset small-range DEM super-resolution model to obtain preset small-range fine DEM data, wherein the method comprises the following steps of:

2. The lunar surface multiscale DEM modeling method based on fused terrain features of claim 1, wherein acquiring the LOLA DEM data and LROC NAC DEM data, and acquiring a plurality of preset high-resolution DEM data sets and a plurality of preset low-resolution DEM data sets based on the LOLA DEM data and the LROC NAC DEM data, comprises:

Wherein the first integer multiple is twice the second integer multiple.

3. The lunar surface multiscale DEM modeling method based on fused terrain features according to claim 1, wherein extracting the terrain feature map dataset corresponding to the plurality of preset high-resolution DEM data using a terrain feature extraction operator, comprises:

4. The lunar surface multiscale DEM modeling method based on fused terrain features of claim 1, wherein constructing a residual fused convolutional neural network comprises:

5. The lunar surface multiscale DEM modeling method based on the fused terrain features of claim 1, wherein training the residual fusion convolutional neural network based on the plurality of preset high-resolution DEM data sets, the plurality of preset low-resolution DEM data sets, and the terrain feature map data sets to obtain a multi-resolution DEM model comprises:

6. The lunar surface multiscale DEM modeling method based on fused terrain features according to claim 5, wherein fusing the terrain feature map dataset into the overall loss function of the residual fusion convolutional neural network, comprises:

7. The lunar surface multiscale DEM modeling method based on the fused terrain features according to claim 1, wherein inputting the plurality of preset high-precision DEM data into the preset small-range DEM super-resolution model to obtain preset small-range fine DEM data further comprises: